1. A Framework for Push-grasping in Clutter
Mehmet R. Dogar
Siddhartha S. Srinivasa
Zheng Yang

2. Outline Tasks in Cluttered Environment The planning Framework
The Action Library
Conclusion
Limitation and Future work

3. Tasks in Cluttered Environment
Grasp the coke can in the clutter!

4. Tasks in Cluttered Environment
Issues:
A. Which to move? In what order?
B. How to deal with uncertainty?
C. With what action?
D. Where to move?
Grasp the coke can in the clutter!

5. The planning Framework
A. Which to move?
1. Attempt to pick the goal object, allowing penetrating into others
2. Identify the penetrated objects and add them on the move list (FindPenetrate)
3. plan to move the object on the move list and iterate

7. The planning Framework
A. Which to move?
Backward Planning (recursive call)
Monotone Planning (avoidVol)

8. The planning Framework
B. How to deal with uncertainty?
Determine the initial uncertainty from perception, the region:
U(0)
Track the uncertainty during manipulation -- the evolution of region U(0) :
v[0,1]
Calculate the volume (Volume):
Volume(o,v)

9. The planning Framework
C. With what action?
The Action Library
Generally an action (a):
Or transit action (Goto)

10. The planning Framework
D. Where to move?
The goal object
The object on the movelist
The NGR (negative goal region): the sum of the volume of space used by all the previous planned actions
The goal region G is
StablePose-NGR

12. The planning Framework
Reconfigure:
a. search the action library
b. generate an action
if failed then continue at a
c. generate transit plan(Goto)
----combine b and c as the plan
d. calculate the action volume of the plan
e. determine the penetrated objects
if none then return the plan
f. update the NGR and avoidVol
g. recursive call of reconfigure of the next object on the list
if done
add the recursive plan before plan
if fail then search again
----if all actions tried then return empty

13. The Action Library Push grasp Sweep parameter ways of search
PG(ph, a, d)
S(ph, d)
ways of search
36 discrete direction, v
lateral offset, l
9 sets of aperture, a
Not specified in the paper
capture region
evolution of uncertainty region
the series of capture region during the push-grasp
1. initial region of object
2. all possible poses in contact with the hand

14. The Action Library
Make sure the initial uncertainty region is in the capture region of the action
During Push-grasp, minimize the penetrated object by searching different values of v, the direction to grasp
Calculate the volumn by sampling from the uncertainty region and perform the Volumn operator
Sweeping uncertainty could be large

15. The Action Library Other actions: Goto:
Search the configuration space of the arm using Constrained Bi-directional RRT planner
Pick up:
As a Push-grasp followed by a Goto

16. Conclusion
A new framework for planning beyond traditional pick- and-place actions
Pushing can manipulate large obstacles
Consevative by considering the uncertainty

17. Limitation and Future work
Sweeping will affect the efficiency because the large uncertainty region results in more object in the movelist, especially for the tight space
Open loop solution brings larger uncertainty without sensor feedback between steps
Might consider dynamic obstacles
Pushing is limited on a large plane surface, what if the object is on a small plane, such as stacked casserole in the fridge
Might thought coordination of two robot arms